Abstract: A new method based on wavelet transform is proposed to extract rotor fault signal energy features automatically. The method is motivated by multi-frequencies analysis and the scaling transform theory is applied to resample the original signal at certain time interval. The re-sampled signal is decomposed into a predefined layer with wavelet transform, and the energy features of the frequency band are acquired. It can eliminate the effect of rotor speed and sampling frequency on the distribution of wavelet analysis frequency band. Energy features of the frequency band have unified physical meanings. In a simulation setup of the ZL-3 multi-function rotor, we simulate 128 samples including 4 kinds of common rotor faults, namely, imbalance, rub-impact, oil whipping and misalignment, extract the samples’ energy features, and construct integrated neural networks to recognize the samples’ faults. The results show validity of the proposed method.

Key words: rotor, wavelet transform, feature extracting, neural networks, fault diagnosis